Abstract: Aiming at the high complexity of gearbox fault vibration data in the variable operating condition environment and the difficulty of extracting fault features, a fault diagnosis method for variable operating condition gearboxes based on wavelet packet decomposition of three-channel data fusion and multiscale residual network is proposed. The method utilizes wavelet packet decomposition reconstruction to fuse the three-channel vibration signals of the gearbox and transforms them into a two-dimensional image using Gram angle and image coding methods. A network structure combining a multi-scale convolutional structure and a residual structure is used to diagnose the faults of gearboxes with variable operating conditions. An efficient channel attention mechanism is introduced to enhance the sensitivity of different features extracted under different scales of convolution, so as to improve the model's characterization ability and classification performance. An efficient channel attention mechanism is introduced to enhance the sensitivity of different features extracted under different scales of convolution, so as to improve the model's characterization ability and classification performance. The experimental results show that the proposed method can achieve a diagnostic accuracy rate of 99.59% under the condition of constant speed and variable load fault data, and 98.58% under the condition of constant load and variable speed fault data, which proves that this method can effectively weaken the influence of variable speed and variable load during operation on fault features.